Challenge
What had to change
The team had a working prototype that impressed in demos but failed in production: it hallucinated, lacked auditability, and ran without per-user cost limits. Clinical leadership would not approve it for ward use.
AI · Healthcare
AI clinical assistant, integrated into the workflow nurses already use.
DrRobot wanted an AI assistant that read patient context safely and shortened a real workflow — without becoming another tab clinicians had to switch into.
Approach
How we got it back under control.
Eval-first
Before any model work, we built a golden test set drawn from real anonymised cases. Every change had to clear the bar.
RAG over private data
Retrieval-augmented generation grounded in the patient record, with row-level access controls. The model could only see what the user could see.
Cost ceilings
Per-user spend caps, hard rate limits, and a fall-back path to a smaller model when the budget was hit.
Outcomes
The measurable result.
Approved
For ward deployment after eval clearance
< 2s
P95 inference latency at the bedside
$/user
Bounded — every feature ships with a cap
What shipped
Core features and controls.
Patient-context-aware assistant
Audit log for every inference
Per-user cost cap
Model-fallback chain
Embedded in the existing nurse workflow
On-prem deployment
Stack
Python
OpenAI / open-source LLM
Postgres + pgvector
Kubernetes
They told us the prototype was not safe to ship before we asked. That conversation saved us a launch.
Let's talk
Tell us about your project.
Send the details through WhatsApp and we'll route it to the right person.
Trusted by founders across healthcare, hospitality and professional services. London HQ · Bilingual EN/AR delivery · NDA-friendly